Wang Haifeng
Baidu Chief Technology Officer
CAAI/ACL Fellow
Machine Translation Enters the Deep Learning Era
The development of machine translation can be traced back to the proposal of machine translation concepts in 1947. Over the course of more than 70 years, machine translation has gone through three stages: rule-based methods, statistical machine learning, and neural network machine translation, entering the deep learning era.
The large-scale industrial application of neural network machine translation faces the demand for high quality, multilingual, and cross-modal industrialization. In terms of translation quality, Baidu has developed a neural network machine translation model that integrates rich features, reducing the omission rate by 80%; a multi-agent joint learning model that breaks through the limitations of single model learning capabilities; achieving first place in international authoritative machine translation evaluations, significantly improving translation quality. In May 2015, Baidu was the first in the world to launch a large-scale neural network machine translation product.
In terms of multilingual translation, addressing challenges such as a large number of languages, uneven corpus distribution, complex deployment, and high implementation difficulty, Baidu has developed a translation model based on shared encoders, breaking through the bottleneck of low-resource language translation, and created a unified framework for multilingual translation, significantly improving deployment efficiency, achieving mutual translation among 203 languages, and supporting 41,006 translation directions.
Regarding machine simultaneous interpretation, addressing the pain point of balancing translation quality and latency, Baidu has created a simultaneous interpretation model based on semantic units, with a translation accuracy exceeding 80% and a time delay of about 3 seconds, achieving translation levels comparable to human interpreters. At the same time, breakthroughs have been made in end-to-end simultaneous interpretation models, achieving cross-modal knowledge sharing through synchronous decoding of speech recognition and machine translation. While Baidu continues to innovate and break through in machine translation technology, it also actively engages in open collaboration, jointly hosting machine simultaneous interpretation seminars with Google and Tsinghua University, and releasing a Chinese-English simultaneous interpretation dataset aimed at real speaking scenarios to promote simultaneous interpretation research.
Translating over 100 billion characters daily, cross-language communication is becoming a reality
Machine translation is one of the AI technologies that Baidu began accumulating and building early on. Since 2010, Baidu has conducted systematic and in-depth research in large-scale industrialized machine translation technology, massive translation knowledge acquisition, multilingual translation, and machine simultaneous interpretation, continuously breaking through and innovating technically, with increasingly rich industrial applications. Baidu translation has formed a complete product matrix, including translation PC version, translation APP, AI simultaneous interpretation, and translation open platform, responding in real-time and accurately to the global massive and diverse translation requests, translating over 100 billion characters daily, a growth of 100,000 times compared to a decade ago.
As of now, Baidu translation has served over 500,000 enterprises and developers, covering more than 30 fields, continuously playing a role in people’s daily lives and work, public services, and scientific research: serving hundreds of important international conferences such as the Service Trade Fair, the Import Expo, and the Global Artificial Intelligence Technology Conference; assisting economic development, helping multinational trade platforms/enterprises reduce costs and increase efficiency; freely opening up translations in the biomedical field, collaborating with epidemic prevention volunteer groups, contributing to global pandemic response, etc.
According to data, the global authoritative consulting agency Gartner released the “Hype Cycle for Natural Language Technologies, 2020”, ranking Baidu as a benchmark organization in neural network machine translation. Baidu is the only domestic unit included in the machine translation field. In December 2020, Gartner mentioned in the report “Market Guide for AI-Enabled Translation Services” that Baidu, due to its outstanding performance in machine translation, strongly entered the list of representative global AI translation service providers.
Baidu will always adhere to technological innovation, promote technological progress, and make greater contributions to industrial upgrading, high-quality social and economic development, and national prosperity.
